EIDOS-alpha

• 

  Thumbnail

• 

  Light Mode - Interface

• 

  Section Interface

• 

  How it works

• 

  Study modes

• 

  Features Section

Built with Google Gemini 3

Inspiration

Ever sat in a lecture, scribbling notes so fast you missed what the professor actually said? Or recorded a 90-minute class only to never watch it again because... who has time? That was us. Every day. We were tired of:

• Choosing between listening and writing
• Messy notes scattered across apps
• Hours of recordings with no efficient way to study from them
• Exam season chaos with disorganized materials

EIDOS came from genuine frustration. We wanted a study buddy that could watch lectures with us, take perfect notes, quiz us, and explain things when we got confused. Less time organizing, more time learning.

What it does

EIDOS (Educational Intelligence & Dynamic Optimization System) turns learning chaos into clarity.

Core Features

• Live Capture: Record lectures, capture frames, transcribe speech in real-time, generate structured AI notes
• AI Course Generator: Input a topic → get complete courses with chapters, lessons, quizzes, and final exams
• Video Lectures: Slide-based presentations with animated whiteboard drawings and human-like AI narration
• Study Sessions: Upload documents, chat with an AI tutor that understands YOUR materials
• AI Tutor: Voice-activated assistant you can interrupt mid-sentence (say "hey" or "wait")
• Certificates: Pass the final exam → download your certificate

How we built it

⚡ Gemini 3 Integration

EIDOS is built entirely on Google Gemini 3. Every AI feature in the platform is powered by the Gemini 3 API via the official @google/generative-ai SDK.

We utilized specific capabilities of the gemini-3-flash-preview model to handle distinct complex tasks:

1. Structured JSON Output (JSON Mode): We enforce strict JSON schemas for generating Courses, Quizzes, and Exams. This ensures that the generated content can be parsed programmatically to create interactive UI elements like progress bars, quiz forms, and certificate validation.

2. Streaming & Real-time Generation: For Notes and Written Lectures, we utilize Gemini's streaming capabilities. Instead of making the user wait for a spinner, content is streamed token-by-token directly to the frontend, creating a responsive "magic" effect.

3. Multimodal Vision Capabilities: In Study Sessions, users can upload PDF documents or images. We pass these directly to Gemini 3 as base64 inline data, allowing the AI to "see" charts, diagrams, and text within uploaded study materials to answer specific questions.

4. Text-to-Speech (TTS): For our Video Lectures, we integrated gemini-2.5-flash-lite to generate natural, human-like narration that syncs with the generated slide content.

Tech Stack

Layer	Technology
Framework	Next.js 14+ (App Router)
Frontend	React 19, TypeScript
Styling	CSS Modules + Design Tokens
Database	MongoDB + Prisma ORM
Auth	Better-Auth
AI Model	Google Gemini 3 Flash Preview
AI SDK	Google Generative AI Node.js SDK
3D	Three.js / React Three Fiber
Math	KaTeX

Key Implementation Details

• Speech Recognition: Web Speech API with auto-restart logic for reliability
• AI Prompts: 100+ lines of carefully crafted instructions for natural-sounding narration
• Streaming: Server-Sent Events for real-time content generation
• Certificates: html2canvas + jsPDF for downloadable PDFs

Challenges we ran into

1. Making AI sound human

   • Early versions sounded robotic
   • Weeks of prompt iteration: contractions, filler words, rhetorical questions, personal touches

2. Speech API reliability

   • Web Speech API stops randomly, throws errors, differs across browsers
   • Built auto-restart logic and extensive error handling

3. AI Tutor interrupt detection

   • Detecting "hey" or "stop" while text-to-speech is playing
   • Canceling speech synthesis and resuming listening without false positives

4. Streaming JSON from Gemini

   • Parsing partial responses while maintaining valid markdown
   • Careful buffer management required

5. Course generation pipeline

   • Multi-step generation: outline → chapters → lessons → quizzes
   • Progress updates and failure handling at each stage

Accomplishments that we're proud of

• Live Capture generates comprehensive notes from 20+ minute recordings
• AI Tutor with voice interruption actually works smoothly
• Full course generation with video lectures and embedded quizzes
• Polished UI with glassmorphism, animations, dark/light themes
• Comprehensive documentation (API, architecture, components, user guides)

What we learned

• Prompt engineering is an art - specificity about format, tone, length matters
• Browser APIs are powerful but inconsistent - expect edge cases
• Streaming beats spinners - users prefer seeing content appear gradually
• TypeScript saves time - caught bugs before they happened
• Good architecture pays off - easy to extend and debug

What's next for EIDOS

• [ ] Advanced analytics and learning streaks
• [ ] Collaborative study workspaces
• [ ] Multiple AI model support
• [ ] Offline mode with sync
• [ ] Spaced repetition flashcards
• [ ] Enhancement

Team Information

Kaleemullah Younas

• Role: Full-Stack AI Engineer - GITHUB

Muhammad Umer

• Role: Web Developer & Devops - GITHUB

Contact

• Primary Contact Email: EMAIL

---

Disclaimer

This is an MVP (Minimum Viable Product) deployed using free-tier services. As a result, the live version may occasionally face 429 errors or rate limits. We plan to upgrade to paid tiers or some proprietary models in the future for a more reliable experience.

Built With

• 2.0
• better-auth
• css
• equation
• google-gemini-2.0-flash
• html2canvas
• jspdf
• katex
• lucide
• modules
• mongodb
• next.js
• prisma
• react-markdown
• react-syntax-highlighter
• reactthreefiber
• three.js
• typescript
• webspeechapi